Limitations

Limitations

Keep these visible in the technical document and the demo.

Limitation	Why it matters	Mitigation / next step
Dataset is synthetic espeak-ng audio	Results don't generalise to human speech	Add real baselines and human re-recordings
Labels are exact by construction, not human-reviewed	No accuracy claim about human annotation	Human review pass with recorded tolerance
Default aligner is approximate and relies on energy valleys	Timing error in real speech; intra-chunk tempo changes are hard to see	Validate WhisperX forced alignment on real audio
WhisperX backend is untested	Models couldn't be downloaded in the build sandbox	Test and re-run the evaluation
Noise sensitivity at 20 dB SNR or lower	False pauses and energy flags	Noise-robust detection; stricter insufficient-evidence rule
Single-factor flaws, three speeches, small counts	Easier than real delivery; wide uncertainty	Larger and more varied evaluation set
Thresholds were set a priori and adjusted once on one speech	Possible mild overfitting to dev	Report dev and held-out separately (done)
One reference per text	Findings depend on the chosen reference	Document the reference; consider multiple references
Reference prior in alignment	Boundaries are pulled toward reference pause positions	Evaluate with an independent aligner
Measurements, not verdicts	A deviation is not proof of a flaw, intent or cause	Explanations name feature, span and rule only
Only four rules over pacing, pauses, pitch contour, energy	No articulation, filler-word or content analysis	Out of scope for now